Prostate cancer ranges widely in its behavior, from slow-growing tumors that may never cause symptoms to aggressive cancers that spread rapidly. Getting the classification right determines whether a patient undergoes active surveillance or immediate treatment with surgery, radiation, or hormone therapy. Two of the most important pathological markers for this decision are the Gleason Grade Group (GGG) -- which reflects how abnormal and structurally disordered the tumor cells are -- and the Ki67 proliferation index, which measures what fraction of cells are actively dividing at any given time.
The Gleason Grade Group system (from 1 to 5, replacing the older Gleason score) predicts tumor behavior precisely: compared to Grade Group 1 cancers, Grade Group 5 cancers have a hazard ratio of 23.9 for recurrence, and their 5-year biochemical recurrence-free survival rate drops to just 26 percent. Meanwhile, Ki67 expression above 10 percent correlates with extracapsular extension, seminal vesicle invasion, and a nearly 12 percent increase in disease-specific death risk for each additional 1 percent of Ki67.
Both markers currently require invasive biopsy and immunohistochemical staining of tissue samples. This introduces risks including bleeding, infection, and pain, and systematic biopsy can miss focal tumors or underestimate grade due to sampling error. If these markers could be predicted accurately from pre-existing multiparametric MRI -- imaging that patients already undergo for staging -- the diagnostic pathway could be made safer and faster.
This study from Chongqing Medical University and National University of Singapore investigated whether radiomics-based machine learning models applied to biparametric MRI (T2-weighted imaging plus diffusion-weighted imaging and ADC maps) could predict Ki67 expression and Gleason Grade Group in 122 prostate cancer patients, testing four different ML algorithms across seven MRI sequence combinations.
Radiomics is the systematic extraction of large numbers of quantitative features from medical images that go far beyond what a radiologist would measure manually. From each MRI sequence, the software calculates first-order features (statistics of pixel intensity values such as mean, variance, and skewness), shape features (geometric properties of the tumor region), and texture features (patterns of intensity variation across the tumor, captured by five different mathematical matrices: GLCM, GLSZM, GLDM, GLRLM, and NGTDM).
To capture multi-scale information, a wavelet transformation was applied to each image before feature extraction. Wavelet analysis decomposes images into different frequency components (representing fine versus coarse texture), producing eight sub-band images per MRI sequence. This dramatically increases the feature space: from a base of 107 features per sequence, wavelet transformation produced 851 features per sequence, for a total of 2,553 radiomic features extracted across all three MRI sequences (T2WI, DWI, and ADC).
Tumor regions were manually segmented by two radiologists with 5 and 10 years of experience in abdominal and urinary imaging respectively, with a senior reviewer resolving any disagreements. Segmentation covered the entire tumor while excluding the urethra and adjacent structures. For multifocal tumors, only the largest lesion was included, ensuring the model trained on the most clinically dominant lesion.
To reduce the 2,553 features to a manageable set without overfitting, the team applied Recursive Feature Elimination (RFE) -- an iterative method that trains a model, ranks features by importance, removes the weakest, and repeats until an optimal set of 20 features is identified. SMOTE (Synthetic Minority Oversampling Technique) was used to correct class imbalance between low-grade and high-grade patient groups, and 5-fold cross-validation validated all results.
The study systematically tested four machine learning algorithms: logistic regression (LR), support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN). Each algorithm was applied to seven MRI input combinations: T2WI alone, DWI alone, ADC alone, T2+DWI, T2+ADC, ADC+DWI, and all three sequences combined (ADC+DWI+T2). This produced 28 models for each prediction task, for a total of 56 models across the Ki67 and GGG prediction tasks.
The dataset split patients into training (80 percent) and testing (20 percent) sets randomly, with the procedure repeated five times and results averaged. Ki67 was treated as a binary variable: high expression (greater than 10 percent) versus low expression (10 percent or less), using the median as the cutoff. GGG was similarly dichotomized into low-grade (GGG 1 and 2) versus high-grade (GGG 3, 4, and 5) using the 2019 ISUP consensus classification.
Performance was evaluated using standard ROC analysis, measuring AUC, sensitivity, and specificity for each model. A separate conjoint analysis tested whether any single model could simultaneously make correct predictions for both Ki67 and GGG in the same patient -- a harder task that requires the model to capture the distinct biological information in each marker simultaneously.
For Ki67 prediction (Task 1), the best model was logistic regression applied to T2 plus ADC sequences (LR_T2+ADC), achieving an AUC of 0.8882 with sensitivity of 0.764 and specificity of 0.866. Logistic regression consistently outperformed the other algorithms for this task. The most important single feature was ADC_wavelet-LHH_firstorder_Maximum, a high-frequency wavelet component of the ADC map capturing intensity extremes in the tumor region. Of the 20 selected features, 11 came from ADC and 9 from T2WI, including 7 first-order, 2 shape, and 11 texture features.
For Gleason Grade Group prediction (Task 2), the best model was a support vector machine applied to T2 plus DWI (SVM_T2+DWI), achieving an AUC of 0.9248 with sensitivity of 0.859 and specificity of 0.784. SVM and logistic regression both substantially outperformed KNN and random forest for this task. The most important feature was DWI_wavelet-HLL_glcm_SumAverage, a DWI texture feature capturing the average co-occurrence of intensity levels across the tumor. Notably, 18 of the 20 selected features came from DWI rather than T2WI, suggesting diffusion restriction carries more grade-related information than T2 signal.
An unexpected finding was that adding all three sequences together (ADC+DWI+T2) actually performed worse than T2+DWI alone for GGG prediction (AUC dropped to 0.40), even lower than individual sequences. The authors hypothesize that high-weight features from different sequences may partially cancel each other out when combined, reducing the model's ability to learn coherent patterns. This highlights that more data is not always better and that sequence selection matters.
Across both tasks, wavelet-derived features dominated the selected feature sets: at least 80 percent of the top 20 features in each model came from wavelet-transformed rather than original images, consistent with prior radiomics studies showing that wavelet features capture tumor heterogeneity more robustly than raw intensity statistics.
Ki67 and GGG showed a weak but statistically significant positive correlation (Spearman r equals 0.382, p less than 0.001). This modest correlation makes biological sense: both measure cancer aggressiveness, but through different lenses -- GGG reflects structural disorder of gland architecture while Ki67 reflects how fast tumor cells are replicating. A patient could have high-grade architecture but relatively slow proliferation, or vice versa.
In the conjoint analysis (predicting both markers correctly in the same patient), the best model was LR applied to ADC plus DWI sequences (LR_ADC+DWI), with a combined accuracy of only 0.6230. This is substantially lower than either individual task AUC, reflecting the challenge of simultaneously predicting two biologically distinct markers with a single feature set. The drop in performance confirms that Ki67 and GGG represent complementary, not redundant, aspects of tumor biology.
Seven of the 20 features were shared between the Ki67 and GGG models in the conjoint analysis, including texture features for SmallAreaLowGrayLevelEmphasis, Maximum, Minimum, ZoneEntropy, and Skewness. These overlapping features represent fundamental aspects of tumor heterogeneity captured by both markers, while the remaining features are marker-specific.
Despite the modest accuracy in joint prediction, the conjoint analysis represents a proof-of-concept for what the authors call multi-variable prediction models: a single radiomics pipeline that could, in the future, simultaneously output multiple pathological indices from one MRI scan, providing more comprehensive preoperative characterization than any single test alone.
The study demonstrates that biparametric MRI radiomics without contrast agents can predict both Ki67 expression and Gleason Grade Group with strong performance, achieving AUC values comparable to or exceeding prior studies that also used dynamic contrast-enhanced (DCE) sequences. This is clinically important because bpMRI avoids gadolinium contrast injection and its associated costs and risks, while providing comparable diagnostic information.
The authors note three main study limitations. First, all tumor segmentations were performed manually, which is labor-intensive and introduces intra-observer variability. Automated or semi-automated segmentation would improve reproducibility for clinical deployment. Second, pathology was based on systematic biopsy rather than targeted MRI-guided biopsy or surgical pathology, which can underestimate tumor grade due to sampling error.
Third and most critically, the study has no external validation cohort: all 122 patients came from a single Chinese hospital. Radiomics features are sensitive to scanner hardware, acquisition protocols, and image processing parameters, making generalizability to other institutions uncertain. The study received a radiomics quality score of 8, comparable to the reported mean for prostate imaging studies, but external validation would significantly increase confidence in the findings.
Future directions include using radical prostatectomy pathology as gold standard (rather than biopsy), incorporating 5-year survival as an outcome measure, and testing the framework for simultaneous prediction of additional markers such as HER-2 or androgen receptor expression. The authors envision a future where a single MRI scan replaces multiple invasive pathology tests in the preoperative prostate cancer workup.